WindFlow

Architects decide where the openings go long before anyone simulates how air will move through the rooms. A CFD study is slow to set up, so it often comes too late.

I wanted to move a window and see the air respond.

The input is a building model with an outdoor margin around it. Openings can be toggled live, and fans, AC units, exhaust and diffusers are body-force regions.

Buoyancy runs on a coupled D3Q7 thermal lattice (Boussinesq, Guo forcing). A nested 2:1 fine simulation box is available for isothermal cases.

The solver is D3Q19 lattice Boltzmann with a regularized BGK collision and Smagorinsky LES. One kernel serves both the CPU and the GPU, using ILGPU with CUDA, OpenCL or a CPU fallback.

Up to a million GPU particles with trails show the flow. Slices, probes and room metrics (air changes per hour, mean speed, stagnant fraction, age of air) sit alongside. The app exports a PDF report, a turntable MP4 and screenshots.

WindFlow is in development: there is no public release and no public link yet. The analytic and self-consistency checks pass, including the orifice-equation check for cross-ventilation (ratio 0.973) and CPU/GPU parity (0.02%).

Validation against measured-room benchmarks is still in progress, so for now I treat the results as comparative rather than validated.